⚡ electrical engineering

Physics-Informed Condition Monitoring of SiC Power Modules

This paper proposes a lightweight, physics-informed condition monitoring framework for sintered-packaged SiC power modules that combines cumulative damage indicators, monotonicity constraints, and heavy-tailed uncertainty estimation to accurately track multi-regime aging behaviors and wirebond liftoff events, significantly outperforming purely data-driven baselines on industrial datasets.

Original authors: Mattia Scarpa, Evgeny Kusmenko, Francesco Toso, Mattia Bruschetta, Ruggero Carli, Simon Achatz

Published 2026-08-11

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mattia Scarpa, Evgeny Kusmenko, Francesco Toso, Mattia Bruschetta, Ruggero Carli, Simon Achatz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Heartbeat of Electric Cars

Imagine the electric car of the future not just as a vehicle, but as a high-speed race car that never stops. To make these cars go fast and charge quickly, engineers use special electronic switches made of a material called Silicon Carbide (SiC). These switches are the "heart" of the car's motor system, turning electricity on and off thousands of times a second. However, just like a human heart, these switches get tired. They face a constant battle against heat and stress, which can cause tiny cracks to form inside them over time. If a switch fails while you are driving, the car could stop working or, worse, catch fire.

To keep these cars safe, engineers need a way to check the "health" of these switches while they are working. This is called "condition monitoring." Traditionally, scientists have tried to guess how tired a switch is by looking at simple patterns, like how much the temperature goes up and down. But this is like trying to guess how old a person is just by looking at their face; it works okay for some people, but it misses the details. The big challenge is that these switches can fail in two very different ways: sometimes they wear out slowly and smoothly, like a tire losing air, and other times they suffer sudden, shocking injuries, like a wire snapping inside. The old methods struggle to handle these sudden surprises. This paper asks a simple but vital question: Can we build a smarter "doctor" for these electronic switches that understands the physics of how they age, so it can spot trouble before it happens, even when the damage is sudden and messy?

The Paper's Story: Teaching a Computer to Feel the Pain

The researchers in this paper, working with experts from Infineon Technologies, decided to build a new kind of health monitor for these Silicon Carbide switches. They focused on a specific type of switch that uses a special "sintered" packaging method. Think of this like gluing two pieces of metal together with a super-strong, silver paste instead of regular solder. This method stops one common type of failure (the glue melting) but leaves another one wide open: the tiny wires that connect the chip can suddenly snap or lift off. When a wire snaps, the switch's behavior changes abruptly, like a car engine suddenly sputtering.

The team realized that old computer models were like students who only memorized the textbook answers. They worked great when the data looked exactly like the training examples, but they got confused when the wires snapped. To fix this, the authors created a "Physics-Informed" framework. Imagine you are teaching a robot to predict how a balloon will pop. Instead of just showing the robot pictures of balloons, you also give it the laws of physics: "Air pressure builds up," "Rubber stretches," and "If a sharp object touches it, it pops instantly."

Here is how their new system works, broken down into three clever tricks:

The "Damage Diary" (Feature Engineering): Instead of just feeding the computer raw numbers like voltage and current, the team gave it a "diary" of the switch's life. They calculated a "Miner's Rule" score, which is a way of adding up every tiny bit of stress the switch has ever felt. It's like a fitness tracker that doesn't just count your steps today, but adds up every step you've ever taken to tell you how worn out your shoes really are. This diary helps the computer understand the history of the switch, not just what is happening right now.
The "No-Backwards-Time" Rule (Monotonicity Constraint): The authors knew that a switch can only get worse, never better. A tire can't un-wear itself. So, they programmed the computer with a strict rule: "The health score can only go down." They didn't force the computer to follow this rule by changing its brain structure; instead, they added a "penalty" to its homework. If the computer guessed that the switch got healthier, it got a bad grade. This kept the predictions realistic and stable, even when the data got messy.
The "Wiggle Room" Output (Heavy-Tailed Distribution): This is the most creative part. Most computer models try to give you one single number as an answer, like "The switch is 85% healthy." But the researchers knew that when a wire snaps, the answer isn't just a number; it's a surprise. So, instead of giving one number, their model gives a "cloud" of possibilities. It says, "I think it's 85% healthy, but there's a chance it could be 70% or 90%." They used a special math shape (called a Student-t distribution) that has "heavy tails," meaning it expects big surprises. This is like a weather forecast that doesn't just say "sunny," but says "sunny, but there's a small chance of a sudden hailstorm." This helps the system stay calm and accurate even when a wire snaps unexpectedly.

What They Found

The team tested their new system on a dataset of 24 real-world switches that were put through extreme heat cycles, simulating years of driving in just a few days. They compared their new "Physics-Informed" system against older, simpler methods that just looked at the raw data without the "damage diary" or the "no-backwards-time" rule.

The results were impressive. The new system reduced the error in predicting the switch's life by about 60% compared to the simplest baseline method. While the old methods stumbled and became very confused when the wires snapped (the "lift-off" events), the new system stayed steady. It was able to predict the health of the switches with much higher accuracy, even when the data was noisy or the switches were behaving strangely.

Crucially, the paper shows that the most important part of this success wasn't making the computer brain more complex or bigger. Instead, it was the "feature engineering"—the act of giving the computer the right kind of information (the damage diary and the physics rules). The authors found that even a simple computer model worked wonders when it was fed these physics-based clues.

The study also confirmed that the system's "cloud of possibilities" was very accurate. When the model said there was a chance of a big error, it was usually right. This means the system doesn't just guess; it knows when it is unsure, which is a vital safety feature for real-world cars.

The Bottom Line

This paper suggests that to keep our electric cars safe and reliable, we shouldn't just rely on raw data or complex AI brains. Instead, we need to teach our computers the basic rules of how these parts break down. By combining real-world physics (like the history of stress and the rule that things only get worse) with smart computer learning, the researchers created a tool that is much better at spotting trouble. It handles the sudden, messy failures that used to confuse older systems, making it a strong candidate for the next generation of safety monitors in electric vehicles. The authors conclude that this approach is not only necessary but also light enough to be installed directly inside the car's computer, ready to watch over the engine every time you hit the road.

Technical Summary: Physics-Informed Condition Monitoring of SiC Power Modules

Problem Statement
Silicon carbide (SiC) power modules are critical components in automotive traction inverters, yet reliable in-field condition monitoring (CM) remains an open challenge. While qualification standards like AQG 324 exist, they do not provide a consolidated approach for estimating the health state of devices during operation. Existing methods face significant limitations: physics-of-failure (PoF) models lack real-time applicability; purely data-driven architectures require large labeled datasets and struggle with generalization; and current physics-informed machine learning (PIML) frameworks are often too computationally demanding for embedded deployment.

A specific challenge addressed in this work is the degradation behavior of SiC modules assembled with sintered silver packaging. Unlike traditional solder-based modules, where forward voltage drop ( $V_{DS}$ ) degrades via smooth, quasi-exponential trajectories driven by solder delamination, sintered modules suppress solder failure. Instead, they exhibit multi-regime degradation profiles characterized by continuous drift in bond-pad metallization and abrupt, non-monotonic perturbations caused by wirebond liftoff events. These sudden jumps in $V_{DS}$ pose new difficulties for standard monitoring approaches.

Methodology
The authors propose a condition monitoring framework that integrates three complementary elements to address the complexity of sintered SiC degradation:

Physics-Informed Feature Engineering:
Rather than relying solely on raw sensor signals (line current and relative $V_{DS}$ ), the input vector is augmented with cumulative damage indicators derived from the module's stress history. These include:
- Cumulative junction temperature swing ( $S_{\Delta T_j}$ ) and mean junction temperature ( $S_{T_j}$ ).
- Cumulative load current integral ( $S_I$ ).
- A Miner-rule damage accumulator ( $D$ ) derived from per-cycle Coffin–Manson estimates.
  These features encode the degradation history in a form directly interpretable by the model, acting as a physics-guided prior at the input level.
Monotonicity Constraint via Gradient Penalty:
To enforce the physical expectation that health degradation is non-decreasing, a soft constraint is applied during training. For stateless architectures (MLP, CMLP), a gradient penalty is imposed on the partial derivatives of the output with respect to the cumulative inputs. For the recurrent architecture (Neural ODE), a temporal penalty is applied to the predicted trajectory. This approach avoids the inference overhead of custom activation functions and remains compatible with simple feed-forward networks.
Probabilistic Output Head:
Instead of producing a single point estimate, the network outputs the parameters of a Student-t distribution (location $\mu$ , scale $\sigma$ , and degrees of freedom $\nu$ ). This heavy-tailed distribution is specifically designed to accommodate the large residuals introduced by wirebond liftoff events, providing calibrated uncertainty quantification that is robust to out-of-distribution variance.

The framework was evaluated using three neural architectures: a standard Multi-Layer Perceptron (MLP), a Contextual MLP (CMLP) with a rolling window of past $V_{DS}$ values, and a Neural ODE (NODE) that carries a hidden state across cycles.

Key Results
The framework was evaluated on an industrial power-cycling dataset provided by Infineon Technologies, comprising 24 devices under four distinct operating conditions. A rigorous four-fold cross-validation protocol was used, stratified by operating condition.

Performance Improvement: The combination of cumulative physics-informed features and monotonicity regularization consistently outperformed purely data-driven baselines. The best-performing configuration (NODE with cumulative features and monotonicity) reduced the Mean Absolute Error (MAE) on the lifetime estimate by approximately 59% compared to the baseline, and up to 70% in specific comparisons against the trivial parameter-free estimator.
Stability: The recurrent NODE architecture demonstrated the lowest cross-fold variance, maintaining stable performance even in validation sets containing modules with wirebond liftoff events.
Uncertainty Calibration: The Student-t output head successfully provided calibrated uncertainty. The inclusion of cumulative features improved coverage rates to near-nominal levels (e.g., ~70.9% for 68% coverage) and reduced the median predictive scale by roughly 37% compared to the baseline. The model correctly identified that heavy tails were necessary, with the degrees of freedom ( $\nu$ ) remaining low (around 6) across all cycles.
Feature Importance: The study found that the feature engineering step was the dominant lever for performance. Once cumulative features were included, the differences between architectures diminished, suggesting that encoding the physical history at the input level is more critical than architectural complexity.

Significance and Claims
The paper claims that integrating physical prior knowledge at the feature, constraint, and output distribution levels is both necessary and sufficient to handle the complexity of SiC power module degradation with sintered packaging.

Key contributions include:

Demonstrating that feature-level physics integration (cumulative stress history) is more effective for generalization than relying solely on architectural complexity or post-hoc constraints.
Showing that the proposed framework remains lightweight enough for embedded deployment while outperforming complex data-driven baselines.
Providing a method that is robust to abrupt failure modes (wirebond liftoff) through heavy-tailed probabilistic outputs, a capability often missing in standard prognostic tools.
Validating that the approach bridges the gap between accelerated test characterization and in-field health estimation, offering a portable solution across different failure mechanisms (as confirmed by a companion study referenced in the conclusion).

The authors conclude that while the current results are promising, future work should focus on broader operating envelopes that mix mission profiles within single trajectories and the development of online adaptation phases for normal converter operation.

The Invisible Heartbeat of Electric Cars

The Paper's Story: Teaching a Computer to Feel the Pain

What They Found

The Bottom Line

Technical Summary: Physics-Informed Condition Monitoring of SiC Power Modules

More like this